A Framework for Evaluating Early-Stage Human - of Marcus Hutter
A Framework for Evaluating Early-Stage Human - of Marcus Hutter
A Framework for Evaluating Early-Stage Human - of Marcus Hutter
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
fer to non-linguistic experience, then in NARS it is possible<br />
to link “Garfield” to related visual images and operation sequences,<br />
so as to enrich its meaning. However, it is important<br />
to understand that both linguistic experience and nonlinguistic<br />
experience are special cases <strong>of</strong> experience, and the<br />
latter is not more “real” than the <strong>for</strong>mer.<br />
In the previous discussions, many people implicitly suppose<br />
that linguistic experience is nothing but “Dictionary-<br />
Go-Round” (Harnad, 1990) or “Chinese Room” (Searle,<br />
1980), and only non-linguistic sensorimotor experience can<br />
give symbols meaning. This is a misconception coming<br />
from traditional semantics, which determines meaning by<br />
referred object, so that an image <strong>of</strong> the object seems to be<br />
closer to the “real thing” than a verbal description. NARS’<br />
experience in Chinese is different from the content <strong>of</strong> a<br />
Chinese-Chinese dictionary, because a dictionary is static,<br />
while the experience <strong>of</strong> a system extends in time, in which<br />
the system gets feedback from its environment as consequences<br />
<strong>of</strong> its actions, i.e., output sentences in Chinese. To<br />
the system, its experience contains all the in<strong>for</strong>mation it can<br />
get from the environment. There<strong>for</strong>e, the system’s processing<br />
is not “purely <strong>for</strong>mal” in the sense that the meaning <strong>of</strong><br />
the symbols can be assigned arbitrarily by an outside observer.<br />
Instead, to the system, the relations among the symbols<br />
are what give them meaning. A more detailed discussion<br />
on this misconception can be found in (Wang, 2007),<br />
and will not be repeated here.<br />
In summary, NARS satisfies the two requirements <strong>of</strong> embodiment<br />
introduced previously:<br />
Working in real world: This requirement is satisfied by<br />
the assumption <strong>of</strong> insufficiency in knowledge and resources.<br />
Having grounded meaning: This requirement is satisfied<br />
by the experience-grounded semantics.<br />
Difference in Embodiment<br />
Of course, to say an implementation <strong>of</strong> NARS running in<br />
a laptop computer is “already embodied”, it does not mean<br />
that it is embodied in exactly the same <strong>for</strong>m as a human mind<br />
operating in a human body. However, here the difference is<br />
not between “disembodied” and “embodied”, but between<br />
different <strong>for</strong>ms <strong>of</strong> embodiment.<br />
As explained previously, every concrete system interacts<br />
with its environment in one or multiple modalities. For a human<br />
being, major modalities include vision, audition, tactile,<br />
etc.; <strong>for</strong> a robot, they include some human-like ones, but also<br />
non-human modalities like ultrasonic; <strong>for</strong> an ordinary computer,<br />
they directly communicate electronically, and also can<br />
have optional modalities like tactile (keyboard and various<br />
pointing devices), audition (microphone), vision (camera),<br />
though they are not used in the same <strong>for</strong>m as in a human<br />
body.<br />
In each modality, the system’s experience is constructed<br />
from certain “primes” or “atoms” that is the smallest units<br />
the system can recognize and distinguish. The system’s processing<br />
<strong>of</strong> its experience is usually carried out on their compound<br />
“patterns” that are much larger in scale, though short<br />
in details. If the patterns are further abstracted, they can<br />
177<br />
even become modality-independent “symbols”. This is the<br />
usual level <strong>of</strong> description <strong>for</strong> linguistic experience, where<br />
the original modality <strong>of</strong> a pattern, with all <strong>of</strong> its modalityspecific<br />
details, is ignored in the processing <strong>of</strong> the message.<br />
However, this treatment does not necessarily make the system<br />
disembodied, because the symbols still comes from the<br />
system’s experience, and can be processed in an experiencedependent<br />
manner.<br />
What makes the traditional symbolic AI system disembodied<br />
is that the symbols are not only abstracted to become<br />
modality-independent, but also experience-independent, in<br />
the sense that the system’s processing <strong>of</strong> the symbol is fully<br />
determined by the system’s design, and have little to do with<br />
its history. In this way, the system’s body becomes completely<br />
irrelevant, even though literally speaking the system<br />
exists in a body all the time.<br />
On the contrary, linguistic experience does not exclude<br />
the body from the picture. For a system that only interact<br />
with its environment in a language, its experience is linguistic<br />
and amodal, in the sense that the relevant modality is not<br />
explicitly marked in the description <strong>of</strong> the system’s experience.<br />
However, what experience the system can get is still<br />
partially determined by the modality that carries out the interaction,<br />
and there<strong>for</strong>e, by the body <strong>of</strong> the system. As far<br />
as the system’s behavior is experience-dependent, it is also<br />
body-dependent, or embodied.<br />
Different bodies give a system different experiences and<br />
behaviors, because they usually have different sensors and<br />
operators, as well as different sensitivity and efficiency on<br />
different patterns in the experience and the behavior. Consequently,<br />
even when they are put into the same environment,<br />
they will have different experience, and there<strong>for</strong>e different<br />
thoughts and behaviors. According to experience-grounded<br />
semantics, the meaning <strong>of</strong> a concept depends on the system’s<br />
experience on the concept, as well as on the possible operations<br />
related to the concept, so any change in the system’s<br />
body will more or less change the system’s mind.<br />
For example, at the current stage, the experience <strong>of</strong><br />
NARS is purely linguistic, so the meaning <strong>of</strong> a concept like<br />
‘Garfield’ only depends on its experienced relations with<br />
other concepts, like ‘cat’, ‘cartoon character’, ‘comic strip’,<br />
‘lazy’, and so on. In the future, if the system’s experience is<br />
extended to include visual and tactile components, the meaning<br />
<strong>of</strong> ‘Garfield’ will include additional relations with patterns<br />
in those modalities, and there<strong>for</strong>e become closer to the<br />
meaning <strong>of</strong> ‘Garfield’ in a typical human mind. There<strong>for</strong>e,<br />
NARS implemented in a laptop and NARS implemented in a<br />
robot will probably associate different meaning to the same<br />
term, even though these meanings may have overlap.<br />
However, it is wrong to say that the concept <strong>of</strong> ‘Garfield’<br />
is meaningful or grounded if and only if it is used by a robot.<br />
There are two common misconceptions on this issue. One is<br />
to only take sensorimotor experience as real, and refuse to<br />
accept linguistic experience; and the other is to take human<br />
experience as the standard to judge the intelligence <strong>of</strong> other<br />
systems. As argued previously, every linguistic experience<br />
must be based on some sensorimotor experience, and though<br />
the latter is omitted in the description, it does not make the<br />
<strong>for</strong>mer less ‘real’ in any sense. Though “behave according